How ETA Brain finds an answer

01The big pictureYour question’s journey

The best way to think about this is as an assembly line. Quality control tidies up the incoming question without changing what it’s asking. Then the work splits between two librarian scouts running in parallel: one hunting exact wording, like an index, the other hunting by topic, more like a map. They come back with a heap of research that might apply. A sorter works through the pile and keeps what it knows the writer will find useful. On the bigger questions, it also checks for gaps in the research. If it finds any, it sends questions back to the librarians for more citations. Finally it lands at copy, where the writer drafts the article from what survived and pins every claim to the exact moment in the episode where somebody said it.

That’s the line, end to end. The rest of this part walks it slower: the paperwork between stations, one real question going through, the one station that isn’t a language model, and the loading dock where episodes become the archive in the first place.

02One real questionWatch a real question go through

Everything above is the machine in the abstract, and diagrams have a way of flattering their subject. So here’s one real question, asked on 2026-07-06 and captured station by station exactly as the pipeline saw it: the ranked pile, the sorter’s cuts, the finished answer. Nothing below is mocked up.

03The strange stationHow a machine matches meaning, not words

The index scout is easy to picture. It matches the exact words you typed, the way a book’s index would. The map scout needs explaining, because its map is not a figure of speech. The scout reads a piece of text and turns it into a point in space: an actual list of 384 numbers (the “meaning fingerprint” from the paperwork above), coordinates on a map with 384 axes instead of two. The map is arranged so that distance means difference in meaning. Two sentences that say the same thing in different words land in almost the same spot. Two that share words but not meaning land far apart.

Every snippet in the archive got its coordinates once, at ingest. Your question gets its coordinates the moment you ask. Then the search is pure geometry, and the database runs it with no model at all: find the snippets sitting nearest your question. Nearest on the map means most likely to answer you, even when a passage shares not one word with what you typed. It’s why a search for “letting a manager go” surfaces a guest talking about “having to fire someone.”

This is the one station on the line that isn’t a language model, which raises a fair question: why not Claude here too? Because Claude is trained to take text in and write text out. It never hands back coordinates, and its internal sense of a sentence was never tuned so that distance equals meaning. Drawing the map is a separate craft with its own training, done by a different kind of model (a dedicated embedder), in our case a small open-source one we run ourselves. Every other AI station on the line is Claude doing a language job. This one earns its place by turning meaning into geometry the database can search in milliseconds.

04The loading dockHow episodes become the archive

None of this works until the episodes are on the shelves, and that happens at the loading dock, long before anyone asks a question. New episodes arrive from each show’s public feed. If the publisher posts a transcript on the episode page, we take it as is. If not, a Mac mini in the workshop listens to the audio and types the transcript itself, using a speech-to-text model (WhisperX) plus a second model that tells the voices apart (the term of art is diarization). It works through one episode at a time and takes about half an hour each.

Then every voice gets a name and a role, host or guest, so an excerpt can say who said it. The audio is deleted the moment the transcript exists. We never host it: when you want to listen, every link on this site sends you to the publisher. We keep short excerpts, they keep the show.

Next the transcript is cut into small units, 2 to 3 sentences each, stored with a wider stretch of the surrounding talk and a timestamp accurate to the sentence. (Part two explains why that size wins.) Each unit gets its coordinates on the meaning map once, right here at the dock. Finally the finished shelves are pushed, one way, to the hosted database the site reads. The working copy never leaves our own machine, and a timer re-checks the published copy every 15 minutes. So far the dock has processed 121 episodes, about 187 hours of conversation, into the 48,577 searchable moments the scouts comb.

That’s the tour: the line, the paperwork, one real question, the strange station, and the dock that feeds them all. Everything above is running right now. What’s left is how the line got this way.

05The original problemThe inspector who demanded two stamps

The sorter’s station used to belong to an inspector with one rigid rule and no judgment. How he lost the job starts on launch day, when the answers were bad in a way that took real effort to pin down: reasonable questions came back empty while outright gibberish sailed through to a paid answer. The culprit was the inspector, standing between the scouts and the writer (technically a relevance gate on the fused RRF score, the merged points from both scouts’ lists). The rule he enforced: a passage moves on only if both scouts stamp it.

The rule sounds sensible until someone misspells a word. The index scout can only match strings that actually appear in the transcripts, so a typo sends him back empty-handed. Once his stamp is missing, the arithmetic is final: no passage can clear the bar, however confidently the map scout stamps it. One wrong letter shut down the entire question.

Here’s the thing: the inspection failed in the opposite direction too. A rescue clause, written to spare near-miss questions from an empty result, waved lone weak hits through whenever the pile came up short. A short pile is exactly what gibberish produces. So nonsense reached paid AI calls while honest questions starved: strict exactly where it should’ve been lenient, lenient exactly where it should’ve been strict.

Replacing him took a day. Choosing the replacement took measurements.

06The fixWe didn’t argue. We held an exam

The exam: 44 tough questions (straightforward, big-picture, typo’d, reworded, pure nonsense) for which we knew in advance exactly which moments in which episodes held the answer. Seven retrieval designs sat it, from carefully tuned classical search to the fashionable techniques in the field’s playbook.

The field deserves naming, because a couple of the entries are the going advice for exactly this problem. There was a cross-encoder, a small model trained for the narrow job of scoring one passage against one query. There was HyDE, where the AI dreams up an ideal answer first and the scouts then hunt for real passages that resemble the dream. There was an episode-first design that ranks whole shows before the moments inside them. And there was the humblest entry on the sheet: keep the launch-day inspector, just rewrite his rules to be reasonable.

The winner, by a wide margin, was none of the fashionable techniques: hire a sorter (the term of art is LLM reranking). Let everything plausible into the pile. Then pay a fast AI to actually read all ~50 candidates and keep the twelve it knows the writer will find useful. A sorter that reads understands the intent behind a typo. It spreads its picks across different episodes. And it’s allowed to conclude that nothing in the pile helps, which is how nonsense finally gets turned away. It found 2½× more of the known-right moments than the launch design.

The most instructive loser was the kitchen-sink entry, which stacked all the clever techniques at once and scored worse than the sorter alone. Turns out improvements don’t add. Each stage deposits its own noise into the pile, and every stage downstream has to fight through it.

07The ladderClimbing, one honest step at a time

With the sorter hired, we kept climbing, under one house rule: every new idea takes the same exam, and it keeps its place only if the numbers improve. We call the rule keep-if-wins, and the result gets written down whichever way it goes. Seven ideas have earned a place so far. Four didn’t, among them some of the field’s favorites and one we had asked for ourselves. The three newest wins went up in July 2026, after we surveyed what other AI search systems were doing and ran the promising ideas up the same ladder.

Two more additions from that run take no exam, because they change no answer. The line now keeps a log of what real visitors ask, so future exams can use real questions instead of only ours. And the answer pages now carry the markup that lets search engines and AI assistants read them properly.

The streaming row hides the run’s best lesson. We predicted first words on screen in 2 to 3 seconds, because we assumed the writing was the slow part. The measurement said no. The stations before the writer, quality control through the sorter, set a floor of about 7 seconds, and streaming only removes the wait for the writing itself. This page argues for measuring instead of trusting a good story. This time the good story was ours.

Why smaller pieces mattered so much

08What we learnedJudgment beats scorecards

The net effect: a typo’d question went from never working to working nine times in ten; answers that once leaned on a single episode now weave together about seven; citations that pointed within a minute of the quote now land within thirteen seconds, and an audit of every claim against its cited excerpts came back 116 for 116; and judges score the answers 7.2 → 8.4 out of ten since the new engine shipped. If someone has already asked your question, the answer arrives instantly. The line built it the first time.

So: ask something big and slightly misspelled. The line is running →

Built July 2026 by Katyella. Every number on this page comes from a frozen, repeatable test (44 golden questions with verified answer locations, plus a 14-question answer eval judged twice per run) and traces to the project’s evaluation records. The full ledger, launch config to today: right moments reaching the writer 15 → 39 of 100; typo survival 0 → 86%; nonsense rejection 0 → 100%; episodes woven into each answer 1.3 → 7.5; citation precision ±1 minute → ±13 seconds; cited claims backed by their excerpts 116 of 116; judged answer quality, measured from the day the new engine shipped, 7.22 → 8.43 of 10. A fresh factual question costs about 3¢ and starts answering in about 7 seconds; a researched big-picture one costs about 6¢ and starts near 20; repeats are free.